Motif 3 MQ87-88-FIT and native ds4 reproduction

This repository records the official-final-only conversion and four-H200 bring-up of the full-topology Motif-3 MQ87-88-FIT GGUF. The immutable mixed weights are public at Baekpica/Motif-3-Mixed-Quant-GGUF, and the preceding full Q8_0 reference is public at Baekpica/Motif-3-GGUF.

The recipe retains all 53 target layers, the first two dense layers, all 51 sparse layers and all 384 routed experts per layer, top-8 routing, the shared expert, Expert-Specific PolyNorm, GDLA, modified mHC, and the complete one-layer MTP block. Nothing is pruned, merged, expert-dropped, layer-dropped, or distilled.

Completed pipeline

87-88 GiB is a nominal capacity class, not a cosmetic hard cutoff. This artifact was accepted at its measured 87.6957 GiB without another precision-reducing pass. On the final native-sm_90 H200 build, the resident model/runtime plus a 262,144-token latent-cache session produced a conservative 95.298828125 GiB CUDA allocation delta on the higher of two resident repeats. Actual GB10 unified-memory residency and OS headroom remain target-host measurements.

Pinned inputs

• Motif 3 final: ccceb1a5fd7b5eb32e47841216b3caf5666c07bc
• Official Motif vLLM fork: 4cd9eb4129883565e69d508038d783d59ee01867
• Official training example: f2cad4115bde073bc61dead4dce8db9711db0ab6
• llama.cpp conversion base: 1d2869c6e54d5003f3927a79efbca0fefa034a6d
• Baekpica/ds4 base: b0309611041655f4e45671cfd9c9886aff161406
• Baekpica/ds4 Motif implementation: feature/motif-3-model-loader@d878ea1a1d67bc0f0bd60e20e75b4a011aa2d8d9
• Mixed-weight revision: efd6044e25e7f8e3b459a737d021091e2e69b6c6
• Q8 reference revision: 5c266c95bf8c8d822d50e5e1cce9d108eaadb2af

Motif-3-Beta was not used as a source, oracle, fixture input, calibration input, or fallback.

Repository map

• manifests/: source hashes, full tensor inventory, locked topology, source maps, quant policy, and verification outputs
• converter/: structural GGUF and source-inventory tooling
• calibration/: corpus construction and real Q8 activation collector; the public tree includes composition but not redistributed corpus text
• fixtures/official-final/: official-equation router, PolyNorm, mHC, GDLA, tokenizer, and chat fixtures
• fixtures/long-context/: exact deterministic 32K/64K/128K/256K native inputs and a 262,080-token decode-reserved 256K OpenAI fixture published with the final H200 bundle
• Native runtime: public Baekpica/ds4:feature/motif-3-model-loader at the pinned implementation commit above
• reports/: artifact, H200 validation, and DGX Spark handoff records
• publish/: public Q8 and mixed model cards plus verification manifests

The private hf://buckets/Baekpica/motif-3-spark-handoff snapshot additionally contains the exact rendered calibration corpus, the 742,004,501-byte imatrix, all four rank-local accumulators, checksum manifest, and Spark pull script. The public GGUF shards are referenced by immutable revision and SHA-256 rather than duplicated in that bucket.

The publication-role guard verifies that these remain two public HF model repositories, one public GitHub reproduction, one public ds4 branch, and one private bucket—and that no HF model repository is created for the reproduction:

bash scripts/audit_publication_layout.sh

Audit and reproduce

Use Python 3.11 or newer. Lightweight manifest/fixture tests run with:

python3 -m pip install -e '.[test]'
pytest -q

Regenerating official tokenizer or long-context fixtures additionally needs the reference dependencies: python3 -m pip install -e '.[reference]'.

The H200 build used /motif3 on Runpod's faster root-backed volume for source weights, GGUFs, imatrix state, and temporary conversion I/O. scripts/env.sh contains the exact default layout. Pipeline entry points are deliberately separate so each irreversible large artifact can be verified before the next stage:

bash scripts/download_sources.sh metadata
bash scripts/download_sources.sh weights
bash scripts/fetch_tensor_headers.sh
bash scripts/build_inventory.sh
bash scripts/build_reference_fixtures.sh
bash scripts/build_tokenizer_fixtures.sh
python3 fixtures/build_server_decode_fixture.py \
  --tokenizer /motif3/source-final \
  --source-text fixtures/long-context/context-262144.txt \
  --source-answer fixtures/long-context/context-262144.answer.json \
  --out fixtures/long-context

Q8 conversion, activation collection, mixed quantization, split, and verification entry points are retained under converter/, calibration/, and scripts/. They require the pinned source snapshot, llama.cpp conversion base, and ds4 quantizer recorded above. The private handoff supplies the exact expensive calibration outputs so Spark work must not requantize the model.

For the native runtime, check out the pinned public ds4 implementation commit. The private handoff additionally preserves ds4/HEAD, status, patch metadata, and complete Motif-touched sources as an offline audit snapshot. Build H200 with the recorded sm_90 configuration. On DGX Spark, rebuild from clean objects with make cuda-spark and confirm the log emits compute_121a/sm_121a; do not reuse H200 binaries.

Scope boundary

The H200 evidence covers strict CUDA weight residency, source-page release, production latent cache allocation/lifecycle, native generation, the OpenAI-compatible server, tools, continuous batching, and completed context gates through 128K. The two partial 256K prefill attempts did not decode and are not correctness passes. This repository does not claim completed single-DGX-Spark 262,144-token serving. That claim requires the immutable artifact to pass the supplied 32K/64K/128K/256K and API gates on one GB10 with model, cache, workspace, server, and required MemAvailable resident, without SSD streaming or CPU weight offload.